Should We Really Edit Language Models? On the Evaluation of Edited Language Models

Model editing has become an increasingly popular alternative for efficiently updating knowledge within language models. Current methods mainly focus on reliability, generalization, and locality, with many methods excelling across these criteria. Some recent works disclose the pitfalls of these editing methods such as knowledge distortion or conflict. However, the general abilities of post-edited language models remain unexplored. In this paper, we perform a comprehensive evaluation on various editing methods and different language models, and have following findings. (1) Existing editing methods lead to inevitable performance deterioration on general benchmarks, indicating that existing editing methods maintain the general abilities of the model within only a few dozen edits. When the number of edits is slightly large, the intrinsic knowledge structure of the model is disrupted or even completely damaged. (2) Instruction-tuned models are more robust to editing, showing less performance drop on general knowledge after editing. (3) Language model with large scale is more resistant to editing compared to small model. (4) The safety of the edited model, is significantly weakened, even for those safety-aligned models. Our findings indicate that current editing methods are only suitable for small-scale knowledge updates within language models, which motivates further research on more practical and reliable editing methods. The details of code and reproduction can be found in https://github.com/lqinfdim/EditingEvaluation.

Paper

Similar papers

Peer review

Reviewer FFCf5/10 · confidence 2/52024-07-12

Summary

The paper explores the general abilities of post-edited language models. Concretely, the paper performs a comprehensive empirical evaluation no various model editing methods and language models. The paper summarizes key findings based on the number of edits, the scale of the language model, the type of tuning, and the safety. The paper shows various numerical metrics to support these findings. The supplementary provides detailed results for experiments and the benchmark construction.

Strengths

1. The paper is well-written, precise, and easy to follow. All findings are summarized and listed. 2. The experiments are comprehensive and detailed. The related dataset, implementations, and evaluation metrics are provided in detail and well-organized. 3. The evaluation includes several latest language models and editing methods which are useful for later research.

Weaknesses

1. Though the experiments are comprehensive and organized, the findings are more about empirical observations rather than systematic analysis. The paper does not distill or form a theory or systematic justification based on these findings. Thus, it is hard to judge academic contributions, especially in terms of the criterion of Neurips. 2. Though the experiments are comprehensive and organized, the findings (line155~159) are more about empirical observations that do not form systematic logic. 3. The paper does not propose any new algorithms to improve the sequential model editing.

Questions

1. Is there any insight to improve the sequential model editing based on these findings? 2. In the experiments of sequential editing, the relation of these edits is not presented. For example, if there are 20 edits, does the correlation of edits in different order influence the conclusion and results?

Rating

5

Confidence

2

Soundness

4

Presentation

4

Contribution

3

Limitations

The author does not provide any limitations for the evaluation method.

Authorsrebuttal2024-08-13

Kindly Reminder

Dear Reviewer FFCf: Thank you very much for your dedicated review. In the rebuttal period, we provide detailed responses to all your comments and questions point-by-point. A brief summary of our responses includes: + Q1: A more systematic and general view of our findings: "Elasticity and Plasticity Trade-off". + Q2: A boarder view of research questions. + Q3: Why are there no new algorithms in our paper. If there are any remaining issues or further clarifications needed, please let us know. We are more than happy to provide additional information or explanations to help resolve any concerns. Thank you for your time and valuable feedback. Best regards

Reviewer FFCf2024-08-14

Response to the rebuttal

Thanks for the clarification and additional results! The results have addressed my questions. I choose to raise my score from 4 to 5.

Reviewer mS2B6/10 · confidence 4/52024-07-13

Summary

The work evaluates the impact of various editing methods on LLMs. Specifically, it investigates how different editing techniques affect the general abilities of models, considering factors such as the number of edits, model scale, safety, and different aspects of model capabilities. Interesting findings and conclusions are suggested by the results.

Strengths

+ The work effectively outlines the problem statement related to editing language models and provides a clear framework for evaluating the impact of different editing methods on model performance. Moreover, the study includes a thorough literature review on LLM knowledge editing. + The empirical studies conducted in the research are solid and rigorous. The study offers reliable findings that can inform future research and development in the field. + Safety is a great perspective to study the current challenges of LLM editing.

Weaknesses

- The outcome of the empirical study is somehow limited and can be further extended. The results are pretty intuitive and do not surprise people much from the current understanding. The authors can provide further insights on the following aspects. 1) Among different editing methods, what are really being traded off when number of editing increases (e.g., does PMET trade editing performance to preserve more general capabilities? And why?). It’d be great to see some hypothesis/quantification to identify and push the boundaries of performance tradeoff. The outcome will be generalizable to future development of methods. 2) Also, given the results, the audience may be further interested in why the muting effect could happen and why instruction-tuned models are more robust. Performing deeper analysis and find associations on the effects would add more value to the results. - The studied problem can be better motivated. If edit needs to be done at scale, e.g., more than 1k edit, is LLM editing still desired or a refreshed fine-tuning could already do the job better? Longer context and retrieval based generation for time sensitive knowledge is also usually consider as common solutions, which may weaken the need for sequential editing at scale. - For methods such as PMET and MEND that are robust within 10^3 editing, could the authors further extend the scale of # edits to verify if the a similar drastic performance drop, as well as the muting effect happens? - The term “intrinsic knowledge structure” appears multiple times to explain the potential reason. The claims can be more solid if a definition and detail discussion can be provided regarding the knowledge structure.

Questions

Discussed in weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

n/a

Reviewer 2vUP6/10 · confidence 4/52024-07-13

Summary

[Post rebuttal update] The authors have addressed my main concern about providing an overarching framework for their evaluation, using the Elasticity - Plasticity tradeoff. I will raise my score from 5 to 6. This paper evaluates the influence of several model editing methods on the models' general capabilities. Their key dimensions of analysis include performance deterioration after several sequential single edits, the robustness of instruction-tuned models, the effect of LLM size and how editing mechanisms affect model safety. The paper stresses that current editing approaches are suitable only for minor updates and highlights the need for more practical and reliable editing techniques. They reveal that while a few edits have minimal impact, extensive edits can lead to significant degradation, including a phenomenon called muting effect, where models tend to produce empty outputs.

Strengths

1. The study provides a thorough assessment of various model editing methods across different language models, offering valuable insights into their effects on general capabilities. 2. By highlighting the pitfalls of current editing methods, such as knowledge distortion and catastrophic forgetting, the paper addresses critical issues that need to be resolved for model editing approaches to work in practical deployments of LLMs. The research emphasizes the impact of editing on the overall performance of language models, including world knowledge, reading comprehension, reasoning, and safety, rather than just on specific tasks. This broader perspective is crucial for understanding the real-world implications of model editing. 3. The study’s exploration of how editing affects the safety of models is particularly valuable, as it addresses a key concern for deploying language models in sensitive applications. 4. The paper claims to be the first to comprehensively evaluate and analyze the general abilities of edited language models. If this claim is true, then it makes it a timely contribution.

Weaknesses

1. The main weakness of the paper is in the exhaustive nature of the evaluation. While the paper evaluates multiple models, the findings may not generalize across all types of language models or specific applications, potentially limiting the broader applicability of the results. It is unfair to ask the authors to work with tons of LLMs but an application stratified view is missing. 2. The paper focuses more on empirical evaluation rather than providing deep technical insights into why certain methods cause performance degradation or safety issues. A more detailed technical analysis could enhance understanding and drive innovation. Note that by technical analysis, I do not mean theoretical analysis, but rather a deeper insight as to what issues are causing these failure modes. Are we not sanitizing the edits well? Are we overfitting? Is there an early stopping criterion that may work here? While pointing out flaws in existing methodologies is useful, since there are many such instances in the LLM world, well grounded research has to provide insights as to why the methodologies are flawed. 3. As an example of (2) above: The observation of the "muting effect," where models produce empty outputs after extensive edits, is concerning. However, the paper does not propose concrete insights as to why this issue might occur, leaving an important problem unaddressed. 4. There are also several different dimensions to consider for an analysis such as this: Scale of the model, prompting conditions or methodologies, whether models benefit from RAGs or not. Since the authors stress practical deployments, they should ensure their results are consistent with practical deployments as well.

Questions

In addition to points above, I wonder what the author's thoughts are on what their best insights are as to why the model performance degrades? How would a practitioner deploying a complex LLM API go about figuring this out? consider cases where they have full models access or just API access.

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

In addition to addressing the weaknesses above, I encourage the authors to think of a more comprehensive evaluation framework to showcase their work. Evaluations of LLMs are themselves brittle -- hence any evaluation needs to have sufficient power to make a solid claim. In the current situation, the paper is indeed well written, but not statistically strong nor comprehensive.

Authorsrebuttal2024-08-13

Kindly Reminder

Dear Reviewer 2vUP: Thank you very much for your valuable comments. In the rebuttal period, we have provided detailed responses to all your comments and questions point-by-point. A brief summary of our responses includes: + Q1: Can our findings generalize to other language models or other application scenarios. + Q2: Distinction between knowledge editing and FT + Q3: Deeper understanding and systematic analysis: Elasticity and Plasticity Trade-off + Q4: Fairness and reliability of our evaluation & Other dimensions for analysis. + Q5: Why performance degradation occurs & Detecting performance degradation with different access levels. Please let us know if any remaining issues or further clarifications are needed. We are more than happy to provide additional information or explanations to help resolve any concerns. Thank you for your time and valuable feedback. Best regards

Reviewer 2vUP2024-08-12

Thank you for your rebuttal

Dear Authors, Thank you for your thorough rebuttal. I will update my review to raise my score. However, I have some lingering questions that you might consider addressing in your paper to provide more clarity and understanding: Is model editing necessary in the age of APIs? If so, who is going to perform model editing at all? In what situations should model editing be even considered aside from FT or RAGs?

Authorsrebuttal2024-08-13

Thanks for response

Dear Reviewer mS2B: We are glad that our responses address some of your concerns. Regarding the question you raised about the use cases for model editing: We believe that in the era of large models and APIs, model editing has become increasingly necessary, especially when resources are limited or when there are few errors that don’t justify full fine-tuning but still need correction. Consider these typical scenarios: 1. If the entity deploying the model is different from the one training it, the deploying entity might lack sufficient computational resources to correct the model's errors through fine-tuning. 2. If the same organization is responsible for both training and deploying the model, but the model only has minor errors that don’t immediately warrant fine-tuning. 3. In situations that require quick responses, such as news updates or emergency events, model editing offers a rapid method of correction without the need for a complete fine-tuning process. 4. In highly sensitive or tightly regulated fields (e.g., finance, healthcare), model editing provides a precise and controlled way to correct specific errors without affecting other parts of the model. Regarding the use cases for RAG (Retrieval-Augmented Generation) and FT (Fine-Tuning): We believe that fine-tuning is suitable for correcting a large number of errors or injecting domain knowledge that the model has not previously learned. This approach can alter the model's overall behavior or style. RAG, on the other hand, is ideal for injecting domain knowledge into a model without fine-tuning, allowing for dynamic knowledge injection, especially for frequently changing information, without impacting the model's overall capabilities. We also welcome new suggestions/comments from you! Best regards, Author of paper 4258

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC